Explainable Text Classification Techniques in Legal Document Review: Locating Rationales without Using Human Annotated Training Text Snippets

Explainability is critical for legal AI adoption: The paper identifies that lawyers often view text classification systems as "black boxes," creating a significant barrier to their use in high-stakes legal discovery work. Making AI decisions transparent and understandable is essential for responsible deployment in legal contexts.

Locating specific text rationales improves attorney trust: The research demonstrates that identifying which specific text snippets within documents triggered a "responsive" classification allows attorneys to easily evaluate whether the model's decisions are correct, addressing a key requirement for responsible AI in legal settings.

Training on snippet-level annotations improves accuracy: The paper's prior research found that models trained on annotated text snippets (rather than entire documents) achieved better precision, suggesting that more granular, explainable approaches can actually improve performance—not just transparency.

Practical explainability without manual annotation is needed: While snippet-level training improves results, manually annotating text snippets during legal review is impractical. The paper addresses this by developing methods to locate rationales automatically, making explainable AI feasible for real-world legal document review without prohibitive manual labor.

Responsible AI requires both accuracy and interpretability: The research emphasizes that responsible AI in legal discovery demands systems that are both accurate and interpretable, recognizing that cost savings from automation mean little if attorneys cannot verify the model's reasoning.